Merge branch 'master' of github.com:sergiooramas/tartarus

Author: Sergio ORAMAS

src/common.py

@@ -5,25 +5,25 @@
 from sklearn.preprocessing import StandardScaler
 
 # Files and extensions
-DATA_DIR = "/homedtic/soramas/text-tartarus/data"
+DATA_DIR = "/mnt/shared/deep_learning/onieto/ismir_tartarus"
 DEFAULT_TRAINED_MODELS_FILE = DATA_DIR+"/trained_models.tsv"
 DEFAULT_MODEL_PREFIX = "model_"
-MODELS_DIR = DATA_DIR+"/models"
-PATCHES_DIR = DATA_DIR+"/patches"
-DATASETS_DIR = DATA_DIR+"/datasets"
-TRAINDATA_DIR = DATA_DIR+"/datasets/train_data"
-FACTORS_DIR = DATA_DIR+"/factors"
-REC_DIR = DATA_DIR+"/playlists"
-HIVE_DIR = DATA_DIR+"/hive"
+MODELS_DIR = DATA_DIR + "/models"
+PATCHES_DIR = DATA_DIR + "/patches"
+DATASETS_DIR = DATA_DIR + "/datasets"
+TRAINDATA_DIR = DATA_DIR + "/datasets/train_data"
+FACTORS_DIR = DATA_DIR + "/factors"
+REC_DIR = DATA_DIR + "/playlists"
+HIVE_DIR = DATA_DIR + "/hive"
 MODEL_EXT = ".json"
 PLOT_EXT = ".png"
 WEIGHTS_EXT = ".h5"
 MAX_N_SCALER = 300000
 
-#create spectrograms folders
-SPECTRO_PATH = DATA_DIR+"/spectrograms/"
-AUDIO_PATH = "/homedtic/soramas/msd/audio/mp3/"
-INDEX_PATH = DATA_DIR+"/index/"
+# create spectrograms folders
+SPECTRO_PATH = DATA_DIR + "/spectrograms/"
+AUDIO_PATH = "/mnt/shared/deep_learning/onieto/msd/msd/mp3/"
+INDEX_PATH = DATA_DIR + "/index/"
 
 
 def ensure_dir(directory):

src/models.py

@@ -437,7 +437,7 @@ def get_model_8(params):
 
     return model
 
-# Metadata  
+# Metadata
 def get_model_81(params):
 
     # metadata
@@ -457,7 +457,7 @@ def get_model_81(params):
 
     x2 = Dropout(params["dropout_factor"])(x2)
     """
-    
+
     dense4 = Dense(output_dim=params["n_out"], init="uniform", activation=params['final_activation'])
     xout = dense4(x2)
     logging.debug("Output CNN: %s" % str(dense4.output_shape))
@@ -605,7 +605,7 @@ def get_model_813(params):
     'cnn' : {
         'dropout_factor' : 0.5,
         'sequence_length' : 500,
-        'embedding_dim' : 300,          
+        'embedding_dim' : 300,
         'filter_sizes' : (2, 3, 4),
         'num_filters' : 150,
         'dropout_prob' : (0.6, 0.7),
@@ -660,7 +660,7 @@ def get_model_82(params):
         #x = flatten(x)
         #logging.debug("Flatten: %s" % str(flatten.output_shape))
         convs.append(x)
-        
+
     if len(params['filter_sizes'])>1:
         merge = Merge(mode='concat')
         out = merge(convs)

src/spectrograms/create_patches.py

@@ -22,6 +22,9 @@
 MAX_N_SCALER=300000
 MSD = False
 
+PATCH_MEAN = -0.0027567206  # Computed from 50k patches
+PATCH_STD = 0.8436051       # Computed from 50k patches
+
 def scale(X, scaler=None, max_N=MAX_N_SCALER):
     shape = X.shape
     X.shape = (shape[0], shape[2] * shape[3])
@@ -79,7 +82,7 @@ def prepare_trainset(dataset_name, set_name, normalize=True, with_factors=True,
             spec = librosa.logamplitude(np.abs(spec) ** 2,ref_power=np.max).T
             for i in range(0,N_SAMPLES):
                 try:
-                    sample = sample_patch(spec,N_FRAMES)                
+                    sample = sample_patch(spec,N_FRAMES)
                     x_dset[k,:,:,:] = sample.reshape(-1,sample.shape[0],sample.shape[1])
                     if with_factors:
                         y_dset[k,:] = factors[t]
@@ -124,6 +127,7 @@ def prepare_trainset(dataset_name, set_name, normalize=True, with_factors=True,
         scaler_file=common.DATASETS_DIR+'/train_data/scaler_%s_%sx%s.pk' % (DATASET_NAME,N_SAMPLES,SECONDS)
         pickle.dump(scaler,open(scaler_file,'wb'))
     return scaler
+  
 
 def prepare_testset(dataset_name):
     spec_folder=common.SPECTRO_PATH+SPECTRO_FOLDER+"/"
@@ -155,5 +159,4 @@ def prepare_testset(dataset_name):
     scaler = prepare_trainset(DATASET_NAME,"train", with_factors=False)
     #scaler = prepare_trainset(DATASET_NAME,"val",scaler=scaler, with_factors=False)
     scaler = prepare_trainset(DATASET_NAME,"test",scaler=scaler, with_factors=False)
-    prepare_testset(DATASET_NAME)
-    
+    prepare_testset(DATASET_NAME)

src/train.py

@@ -216,17 +216,18 @@ def load_data_hf5_memory(params,val_percent, test_percent, y_path, id2gt, X_meta
         index_train = f["index"][:]
         index_train = np.delete(index_train, np.where(index_train == ""))
         N_train = index_train.shape[0]
-        
+
         val_hdf5_file = common.PATCHES_DIR+"/patches_val_%s_%sx%s.hdf5" % (params['dataset']['dataset'],params['dataset']['npatches'],params['dataset']['window'])
         f_val = h5py.File(val_hdf5_file,"r")
         X_val = f_val['features'][:]
         #Y_val = f_val['targets'][:]
         factors_val = np.load(common.DATASETS_DIR+'/item_factors_val_'+y_path+'.npy')
         index_factors_val = open(common.DATASETS_DIR+'/items_index_val_'+params['dataset']['dataset']+'.tsv').read().splitlines()
-        id2gt_val = dict((index,factor) for (index,factor) in zip(index_factors_val,factors_val))            
+        id2gt_val = dict((index,factor) for (index,factor) in zip(index_factors_val,factors_val))
         index_val = f_val['index'][:]
         X_val = np.delete(X_val, np.where(index_val == ""), axis=0)
         index_val = np.delete(index_val, np.where(index_val == ""))                
+
         Y_val = np.asarray([id2gt_val[id] for id in index_val])
 
         test_hdf5_file = common.PATCHES_DIR+"/patches_test_%s_%sx%s.hdf5" % (params['dataset']['dataset'],params['dataset']['npatches'],params['dataset']['window'])
@@ -235,10 +236,11 @@ def load_data_hf5_memory(params,val_percent, test_percent, y_path, id2gt, X_meta
         #Y_test = f_test['targets'][:]
         factors_test = np.load(common.DATASETS_DIR+'/item_factors_test_'+y_path+'.npy')
         index_factors_test = open(common.DATASETS_DIR+'/items_index_test_'+params['dataset']['dataset']+'.tsv').read().splitlines()
-        id2gt_test = dict((index,factor) for (index,factor) in zip(index_factors_test,factors_test))            
+        id2gt_test = dict((index,factor) for (index,factor) in zip(index_factors_test,factors_test))
         index_test = f_test['index'][:]
         X_test = np.delete(X_test, np.where(index_test == ""), axis=0)
         index_test = np.delete(index_test, np.where(index_test == ""))                
+
         Y_test = np.asarray([id2gt_test[id] for id in index_test])
     else:
         hdf5_file = common.PATCHES_DIR+"/patches_train_%s_%sx%s.hdf5" % (params['dataset']['dataset'],params['dataset']['npatches'],params['dataset']['window'])
@@ -271,7 +273,9 @@ def load_data_hf5_memory(params,val_percent, test_percent, y_path, id2gt, X_meta
         X_test = [X_test,X_meta[N_train+N_val:N]]
     return X_val, Y_val, X_test, Y_test, N_train
 
-def batch_block_generator(params, y_path, N_train, id2gt, X_meta = None, val_from_file = False):
+
+def batch_block_generator(params, y_path, N_train, id2gt, X_meta=None,
+                          val_from_file=False):
     hdf5_file = common.PATCHES_DIR+"/patches_train_%s_%sx%s.hdf5" % (params['dataset']['dataset'],params['dataset']['npatches'],params['dataset']['window'])
     f = h5py.File(hdf5_file,"r")
     block_step = 50000
@@ -281,25 +285,25 @@ def batch_block_generator(params, y_path, N_train, id2gt, X_meta = None, val_fro
     if X_meta != None:
         with_meta = True
     while 1:
-        for i in range(0,N_train,block_step):
-            x_block = f['features'][i:min(N_train,i+block_step)]
-            index_block = f['index'][i:min(N_train,i+block_step)]
+        for i in range(0, N_train, block_step):
+            x_block = f['features'][i:min(N_train, i+block_step)]
+            index_block = f['index'][i:min(N_train, i+block_step)]
             #y_block = f['targets'][i:min(N_train,i+block_step)]
             x_block = np.delete(x_block, np.where(index_block == ""), axis=0)
             index_block = np.delete(index_block, np.where(index_block == ""))
             y_block = np.asarray([id2gt[id] for id in index_block])
-            if params['training']['normalize_y'] == True:
-                normalize(y_block,copy=False)
+            if params['training']['normalize_y']:
+                normalize(y_block, copy=False)
             items_list = range(x_block.shape[0])
             if randomize:
                 random.shuffle(items_list)
-            for j in range(0,len(items_list),batch_size):
+            for j in range(0, len(items_list), batch_size):
                 if j+batch_size <= x_block.shape[0]:
                     items_in_batch = items_list[j:j+batch_size]
                     x_batch = x_block[items_in_batch]
                     y_batch = y_block[items_in_batch]
                     if with_meta:
-                        x_batch = [x_batch,X_meta[items_in_batch]]
+                        x_batch = [x_batch, X_meta[items_in_batch]]
                     yield (x_batch, y_batch)
 
 def process(params,with_predict=True,with_eval=True):
@@ -365,7 +369,7 @@ def process(params,with_predict=True,with_eval=True):
             id2gt = dict()
             factors = np.load(common.DATASETS_DIR+'/item_factors_train_'+config.y_path+'.npy')
             index_factors = open(common.DATASETS_DIR+'/items_index_train_'+params['dataset']['dataset']+'.tsv').read().splitlines()
-            id2gt = dict((index,factor) for (index,factor) in zip(index_factors,factors))            
+            id2gt = dict((index,factor) for (index,factor) in zip(index_factors,factors))
             X_val, Y_val, X_test, Y_test, N_train = load_data_hf5_memory(params,config.training_params["validation"],config.training_params["test"],config.y_path,id2gt,X_meta,config.training_params["val_from_file"])
             if params['dataset']['nsamples'] != 'all':
                 N_train = min(N_train,params['dataset']['nsamples'])
@@ -382,7 +386,7 @@ def process(params,with_predict=True,with_eval=True):
     else:
         monitor_metric = 'val_loss'
     early_stopping = EarlyStopping(monitor=monitor_metric, patience=4)
-    
+
     if only_metadata:
         epochs = model.fit(X_train, Y_train,
                   batch_size=config.training_params["n_minibatch"],
@@ -404,7 +408,7 @@ def process(params,with_predict=True,with_eval=True):
                       batch_size=config.training_params["n_minibatch"],
                       shuffle='batch',
                       nb_epoch=config.training_params["n_epochs"],
-                      verbose=2, 
+                      verbose=2,
                       validation_data=(X_val, Y_val),
                       callbacks=[early_stopping])